A Novel Two-Stage Net for Human Body Region Segmentation via Centimeter-Wave Radar

Human body region segmentation is a sub-task in semantic segmentation. Due to the size, complexity, abstraction of radar signal and the lack of open source datasets of radar signal data, the radar signal is rarely used for human body segmentation in past. In recent years, with the development of deep learning, some radar signal data from unmanned drive has been used for research on classification, target detection and semantic segmentation. In this paper, a novel two-stage neural network model based on centimeter-wave radar sensor is designed for human body region segmentation. In the first stage, radar signals are encoded and compressed, in the second stage, decode the spatial feature information of the human body extracted by the encoder to output the segmentation results. The network adopts the design idea of teacher-student network, which use the teacher network to annotate radar signal data and provide cross-modal supervision for the student network. We collected human body data in different scenes, and input the radar signal into the model, performs human body region segmentation. The effectiveness of this human segmentation model is experimentally verified for single and multi-person scenes, especially for occlusion scenes.